TL;DR
- Postmark delivers parsed inbound messages to a webhook and exposes inbound message status. Check its exact acknowledgement and retry rules.
- Twilio SendGrid parses incoming email and attachments into a webhook request. Its documented message-size and retry limits affect recovery design.
- Mailgun supports inbound routes, forwarding and temporary storage. Define where durable message and attachment records live.
- Parseur focuses on extracting business fields from emails and documents. Its document and page credits are different from transport-message charges.
- Parsing email structure is not the same as deciding whether a sales reply is interested or extracting a reliable purchase order. Build and test that interpretation layer separately.
Do you need email parsing or business-field extraction?
Inbound email is a message format, not a tidy CRM record. A single message can contain plain text, HTML, quoted conversations, forwarded messages and multiple attachments. Parsing breaks that message into components such as sender, recipients, subject, bodies, headers and files. Business-field extraction attempts to interpret those components, for example, identifying an order number or classifying a reply as a meeting request.
Choose the layer you need before comparing prices. A transport provider can receive and parse MIME, then send the structured message to your endpoint. Your application may still need to identify the newest reply, extract a field, review uncertainty and write to the CRM. A document parser may provide more of the extraction workflow, but its output still needs a validation and recovery policy.
The four options below are described from provider documentation. We have not tested them on a shared message corpus. For the next step after parsing, our sales reply classification guide covers interpreting responses rather than receiving the raw message.
Inbound email parsing options at a glance
| Option | Starting fit | Output and recovery question | Cost boundary |
|---|---|---|---|
| Postmark | Structured inbound message webhook | Which body, headers and attachments are delivered, and how is a failure replayed? | Inbound plan and volume, storage and processing |
| SendGrid | Inbound Parse for a configured receiving domain | Does the webhook handle payload size and the documented retry window? | Eligible plan, message volume and downstream processing |
| Mailgun | Route and process inbound mail | Is temporary storage sufficient while your application retrieves files? | Routes, messages, storage and downstream processing |
| Parseur | Extract fields from email and documents | How are uncertain fields, templates and attached pages handled? | Documents and pages, not just incoming messages |
A low cost per message is incomplete if the receiving endpoint loses events or the extracted fields require extensive manual correction.
1. Postmark: for structured inbound webhooks
Postmark's inbound webhook delivers parsed message data to an endpoint associated with an inbound message stream. Its payload includes message metadata and attachments; review the actual schema and size limits for your workload.
Postmark's inbound status guidance distinguishes processed, scheduled and failed messages. Its webhook documentation says a non-200 response generally triggers retries, but a 403 stops retries; failed messages can be manually retried. Build monitoring around the inbound stream and check the current retry and replay rules before relying on them. Choose failure responses deliberately so a transient problem is not mistaken for an instruction to stop delivery attempts.
The documented schedule contains ten retries, starting after one minute and ending with a six-hour interval after the preceding attempt. An operator should alert on the first failures rather than waiting for the final retry. Postmark can also provide StrippedTextReply when it can extract reply text, but that field is not guaranteed to isolate every reply correctly; retain the original body and check difficult threads in the fixture pack.
Postmark is a candidate when a team wants structured inbound delivery alongside a transactional email service. Your application still owns durable downstream processing, CRM field mapping and the decision about quoted text or extracted intent. Return the documented success response only after the message has been safely accepted into your processing path.
2. SendGrid: for inbound parse webhooks
Twilio SendGrid's Inbound Parse receives mail for a configured domain or subdomain and posts parsed content and attachments to a specified URL. Its documentation lists a 30 MB total message limit, including attachments, and says the endpoint will not follow redirects.
The developer documentation explicitly says it retries responses with a 5xx status and says messages undelivered after three days are dropped without a prior notification. A support article describes retries more broadly for non-2xx responses over 72 hours. Because those descriptions differ on the exact failure status, confirm the current contract and test your configured endpoint rather than assuming a 4xx response will be retried. Alert an operator before the delivery window expires.
SendGrid fits an application that needs inbound messages delivered to its own processing service and can manage domain setup. It does not by itself decide whether a reply is positive or extract a purchase order reliably. Protect the endpoint and persist the message before acknowledging successful receipt.
3. Mailgun: for inbound routes and temporary storage
Mailgun routes match incoming messages and perform actions such as forwarding or storing. The route-action documentation describes temporary storage for up to three days, which can be useful for later retrieval of a large message or attachment.
Temporary storage is not a long-term archive. Decide where the application stores the message, attachment and processing state, and how it detects a retrieval failure before the provider's retention window closes. Verify the signature and retry behaviour for the specific route delivery, not a different class of outbound event webhook.
Mailgun evaluates lower-priority routes even after a match unless a stop() action ends the route sequence. If two routes forward the same message to the same workflow, both can produce delivery attempts. Design the routing rules and deduplication together; do not assume the first match automatically ends evaluation.
Mailgun is a candidate when a team wants flexible recipient-based routing and already operates an application pipeline. It is less complete if the buyer expects business fields and a reviewed CRM record without building extraction and exception handling.
4. Parseur: for business-field extraction
Parseur focuses on extracting structured information from emails and documents. That can be useful when the buyer's question is “what is the purchase order number and amount?” rather than only “what are the message's MIME parts?” Verify how the chosen extraction method handles the documents and layouts you actually receive.
Its pricing FAQ states that an email and its PDF attachment are treated as separate documents, and that page-based credits apply to processing. One incoming email with a multi-page attachment can therefore represent more than one unit of work. Confirm current plan allowances, page counts, reprocessing rules and export or webhook capabilities.
The same FAQ says credits are used the first time a document is successfully processed, so changing extraction instructions and reprocessing that existing document need not consume another credit under the described rules. Deleting and uploading it again is different: the FAQ says that counts credits twice. It also says documents that exceed the quota are marked for later processing. Confirm those rules for your plan and distinguish an unprocessed quota exception from a successfully extracted document.
Parseur is a candidate when business-field extraction and review are central. It should not be assumed to replace every email transport, mailbox or CRM function. Ask how low-confidence fields are flagged, how templates are maintained when a sender changes format and what evidence a reviewer can inspect.
What does a 100,000-message workflow cost?
Assume 100,000 inbound emails per month, with 10% containing an attachment: 10,000 attachment-bearing messages. The attachment count does not tell you the number of documents or pages. Suppose each such message has one two-page PDF; the workload then contains 100,000 email documents plus 10,000 PDFs and 20,000 PDF pages. These are illustrative inputs, not observed traffic.
For a synthetic budget, assume inbound transport costs $300 per month, storage and monitoring $120, business extraction $400 and ten hours of exception handling at $60 per hour. The total is $1,420, or $0.0142 per inbound message. It excludes initial implementation, tax and any separate model or CRM charges. None of these line items is a quote from a named provider.
For a page-based extractor, ask whether the email body and PDF pages both consume credits. Parseur's published pricing counts a processed email as one credit and a PDF by its pages. If all 100,000 emails and all 10,000 two-page PDFs are successfully processed as separate documents, the workload would require 120,000 credits under those rules. This is an allowance calculation, not a Parseur price quote; actual extraction choices, failed documents and the selected plan affect the bill.
Now include failure handling. If 1% of messages require five minutes of human review, that is 1,000 × 5 minutes = 83.3 hours. At an assumed $60 per hour, review alone costs about $5,000, far more than the ten-hour placeholder above. Do not add both estimates together: the larger figure is a sensitivity scenario replacing the original review assumption. Measure exception rate and review time during a pilot.
A quote should specify message and page volume, attachment size, retention, retries, data region and the extraction or classification layer. The webhook ingestion guide covers the broader delivery and processing architecture.
How do you test the parser and recovery path?
Create a sanitised fixture pack, using synthetic content and files, before connecting real customer mail. Include a plain-text reply, HTML-only message, forwarded chain, quoted thread, non-English characters, multiple recipients, duplicate Message-ID, two attachments, a malformed file and a message above the provider's documented size limit. This is a proposed fixture, not a benchmark we have run.
For each message, record the expected sender and recipients, subject, newest reply, quoted text, attachment names and types, and the business fields the application is allowed to extract. Compare the provider payload with that expectation. MIME parsing may preserve all the text while a separate reply-extraction rule incorrectly treats an old quoted sentence as the new response.
Test failure deliberately. Persist the inbound payload and a stable processing identifier before returning the provider's documented success response. Then simulate a worker crash, duplicate webhook, temporary CRM outage and a retry after a partial write. Ensure one message cannot create two CRM activities or trigger two automated replies. If durable acceptance fails, return the provider-appropriate failure response and alert an operator before its retry window expires.
Protect the processing path. Verify the authentication or signature mechanism documented for that inbound product; do not assume a different webhook's signing method applies. Treat message content and attachments as untrusted, enforce size and file-type rules, scan files where appropriate and avoid logging sensitive bodies unnecessarily. A prompt inside an email should not grant an AI agent new permissions.
For extracted fields, retain the source message, extraction version and confidence or review status. Route missing or conflicting values to a person instead of silently writing a guessed amount to the CRM. Agree retention and deletion rules with the data owner. The GTM engineering playbook helps assign responsibility for that exception queue.
How should an operator handle each failure?
Use a recovery runbook that separates receipt from interpretation. The following is an example policy to test with the provider's actual delivery contract:
| Situation | Record and first action | Recovery decision |
|---|---|---|
| Receiver is unavailable before durable acceptance | Alert on failed delivery and retain provider event ID if available | Follow the provider's retry and replay rules before its retention window closes |
| Same delivery arrives again | Match the provider event ID and stored processing state | Acknowledge an already accepted delivery without creating another CRM activity |
| Message accepted, CRM unavailable | Keep the durable message reference and failed work item | Retry the CRM write under a stable operation key; reconcile an uncertain partial write first |
| Attachment cannot be retrieved or parsed | Retain the message reference, file metadata and failure reason | Retry retrieval while available; route unsupported or damaged files for review |
| Extraction completes with a missing or conflicting amount | Preserve the source and extraction version | Hold the field for review; do not retry transport or silently write a guessed value |
A repeated event and a new representation of a message are not always the same thing. Agree how to identify each with the supplier, and include that distinction in the fixture pack. Set an owner and deadline for the exception queue: an operator cannot recover a message from temporary storage after its actual retention window has expired.
Which inbound email parser should you choose?
Evaluate Postmark for structured inbound webhooks, SendGrid for its inbound parse workflow, Mailgun for configurable routes and temporary retrieval, and Parseur when extracting business fields from messages and documents is the core job. Run the same sanitised fixtures and compare the complete processing and recovery cost.
If the application also needs an agent-owned email identity, the AI agent email API guide covers that adjacent requirement. A parsed inbound event is not automatically a durable agent inbox.
FAQ
Does parsing an email classify the sender's intent?
Not necessarily. MIME parsing separates message components. Intent classification is an additional interpretation step that needs its own categories, confidence rules and human review for uncertain cases.
Why can one email consume more than one extraction unit?
An email can contain several attachments and many document pages. Providers may count the email, each file or each page separately. Confirm the unit and reprocessing rules in the actual plan.
What should happen if my webhook is unavailable?
Follow the provider's current acknowledgement and retry contract, monitor failures and have a replay or recovery process before the retention window expires. Once your endpoint acknowledges receipt, your own durable queue and downstream monitoring become critical.





